Persist ONE durable fact to long-term memory (write-through).

Call this the moment a fact worth surviving compaction appears — do NOT wait for the conversation to be summarized. One call = one atomic fact.

category is domain.sub. Built-in taxonomy: code.rules conventions, constraints, do/don't agreed this session code.workflow current procedure/steps, what's done, what's pending code.os OS, shell, tool versions, paths, env vars code.connections hosts / SSH / ports / domains / services / DBs in use code.files files changed, with absolute paths code.issues unresolved bugs / blockers business.goal the business problem being solved, expected outcome business.decision business decisions + rationale business.constraint requirements, limits, deadlines, stakeholders business.state where we are in the business flow

project: optional slug to scope the fact to one project/repo. tags: optional keywords for later filtering. source: optional origin note (e.g. a file path or URL).

Retrieve relevant facts by meaning (semantic search).

Call this at the start of a task instead of relying on the compaction summary. Narrow with category (code.connections, or a whole domain like code) and/or project. Returns the top matches with a similarity score.

List the active category taxonomy (domains and their sub-categories).

Delete a stored fact by its id (as returned by memory_store/memory_find).

Create the FIRST STATE for a project and wire Claude Code to use memory.

Call this once at the start of working on a repo (or to refresh — it's idempotent). It:

1. Scans the repo into a categorized project.* baseline (overview, stack, structure, commands, connections, git, docs).

2. Folds in current context: ingests any existing compaction summaries for this repo's Claude Code session(s).

3. Installs a managed memory block in the project CLAUDE.md so the agent knows to query (memory_find) and rely on automatic updates.

root: repo path (defaults to the server's working directory). When running in Docker, mount the repo and pass its in-container path here. install_hooks: also add SessionStart/PostCompact hooks to ~/.claude/settings.json.

Capture Claude Code's OWN compaction summaries from disk into memory.

This is the primary write path: instead of tagging facts by hand, it reads the transcript(s) Claude Code writes to ~/.claude/projects/<slug>/*.jsonl, finds every compaction summary, and stores each of the summary's numbered sections (Primary Request, Files and Code Sections, Errors and fixes, Pending Tasks, ...) as categorized, dedup'd chunks. Safe to run repeatedly — unchanged chunks re-map to the same id, so nothing piles up and nothing is lost across compaction generations.

project: restrict to one project slug (the transcript folder name). Omit to scan all projects. session_path: ingest a single .jsonl transcript instead of scanning.

Report collection size, storage backend, and embedding configuration.